SeaLLMs
Open language models for Southeast Asian languages
SeaLLMs are open-weight large language models built by researchers to support Southeast Asian languages. Based on Llama-2, they use continued pre-training and alignment tuning to handle regional linguistic and cultural contexts better than general models.
- Built on Llama-2 with continued pre-training for SEA languages
- Supports Thai, Vietnamese, Indonesian, Burmese and more
- Outperforms ChatGPT-3.5 in several non-Latin languages
- Open weights available for 7B and 1.5B parameter models
- Enhanced safety and reduced hallucination in v3 series
What it is
SeaLLMs are a series of large language models designed to address the linguistic bias in artificial intelligence that often favours high-resource languages like English. Developed by researchers at Sea AI Lab, these models focus specifically on Southeast Asian languages. The family includes versions with 7 billion and 13 billion parameters. They are built to understand and generate text in languages such as Thai, Vietnamese, Indonesian, and Burmese, while also supporting English and Chinese. The latest iteration, SeaLLMs-v3, aims to provide state-of-the-art performance for its size class across tasks like translation, reasoning, and instruction following.
How it came about
The project was introduced in December 2023 through a technical report submitted to arXiv. The work was later presented at the ACL 2024 Demo Track. The initial release focused on correcting the imbalance where regional languages were poorly served by existing global models. The team released subsequent versions, including SeaLLMs-v3, which improved upon earlier capabilities. This third version introduced specific enhancements for trustworthiness and reduced hallucinations, particularly in queries related to local Southeast Asian culture and contexts.
Inside the model
The models are built upon the Llama-2 architecture, a dense decoder transformer. The developers applied continued pre-training to adapt the base model to regional languages. This process involved expanding the vocabulary to include tokens specific to Southeast Asian scripts and linguistic structures. Specialised instruction tuning and alignment processes were used to help the model respect local cultural norms, customs, and legal considerations. The v3 series includes fine-tuning for chat applications, enabling effective multi-turn conversations and better adherence to human instructions.
What it can do
Evaluations show strong performance in multilingual world knowledge, tested using the M3Exam dataset which contains local exam questions. SeaLLMs-v3-7B-Chat achieved high scores in Indonesian, Thai, and Vietnamese compared to similar open-source models. It also performs well in mathematical reasoning using the MGSM dataset. In instruction-following benchmarks like SeaBench, the model demonstrates robust capability in multi-turn settings. The developers claim it outperforms some larger models in non-Latin languages such as Khmer, Lao, and Burmese, while remaining lightweight enough for cost-effective operation.
How to use it
The model weights, code, and demonstrations are released openly. Users can download the models from platforms like Hugging Face. The series includes a 7 billion parameter chat model and a smaller 1.5 billion parameter variant for devices with limited computational resources. Developers are expected to comply with the SeaLLMs Terms of Use. The licence allows for broad usage but requires users to perform their own safety checks and red-teaming before deployment in production environments.
Safety and testing
The developers emphasise trustworthiness in the v3 series, noting efforts to reduce hallucinations and improve safety in responses. They conducted red-teaming and safety fine-tuning to address potential risks. However, they explicitly state that the models may still generate inaccurate or misleading content. Users are warned that the models might produce harmful outputs despite these measures. The team advises stakeholders to implement additional security protocols and adhere to local governance regulations when using the software.
Reception and debate
The project has been recognised in academic circles, with its technical report accepted for the ACL 2024 Demo Track. It is cited as an innovative approach to addressing linguistic bias in large language models. The community has engaged with the release through platforms like Hugging Face, where inference code and evaluation metrics are shared. The models are positioned as a cost-effective alternative for serving Southeast Asian markets, offering performance that rivals or exceeds other open-source options in regional languages.
Why it matters
SeaLLMs contribute to the democratisation of artificial intelligence by providing high-quality tools for low-resource languages. They enable developers in Southeast Asia to build applications that accurately reflect local linguistic nuances. By building on open architectures like Llama-2, the project encourages further research and adaptation in the region. The release of benchmarks like SeaBench helps standardise evaluation for multilingual instruction-following, fostering more inclusive development in natural language processing.
Questions people ask about SeaLLMs
What languages does SeaLLMs support?
It supports many Southeast Asian languages including Indonesian, Vietnamese, Thai, Tagalog, Malay, Burmese, Khmer, Lao, Tamil, and Javanese. It also handles English and Chinese effectively.
Is SeaLLMs free to use?
Yes, the weights and code are released openly. You can download them from Hugging Face. However, you must agree to the Terms of Use and perform your own safety checks before deployment.
How does it compare to other models?
It achieves state-of-the-art performance among open-source models of similar size for Southeast Asian tasks. It reportedly outperforms ChatGPT-3.5 in non-Latin languages like Thai and Khmer.
Good reading and sources
There is no English Wikipedia article on SeaLLMs yet, so the facts on this page come from the sources listed above: huggingface.co/SeaLLMs/SeaLLMs-v3-7B-Chat, arxiv.org/abs/2312.00738. Some are the subject's own pages, so its claims about itself are reported as its claims. EveryLLM wrote the text in its own words, with AI assistance, and compares it with the sources to make sure no passage is copied.
It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.